Fuzzy classification rules based on similarity

The paper deals with the aggregation of classification rules by means of fuzzy integrals, in particular with the fuzzy measures employed in that aggregation. It points out that the kinds of fuzzy measures commonly encountered in this context do not take into account the diversity of classification rules. As a remedy, a new kind of fuzzy measures is proposed, called similarity-aware measures, and several useful properties of such measures are proven. Finally, results of extensive experiments on a number of benchmark datasets are reported, in which a particular similarity-aware measure was applied to a combination of Choquet or Sugeno integrals with three different ways of creating ensembles of classification rules. In the experiments, the new measure was compared with the traditional Sugeno λ-measure, to which it was clearly superior.